Business OS

Inspiration

Launching a small business is not one task. It requires market research, idea validation, supplier discovery, planning, budgeting, risk assessment, and coordinated execution. These activities are usually fragmented across different tools, documents, and conversations.

Business OS was created to turn a business intention into a controlled and executable launch process rather than another isolated AI answer.

What it does

Business OS accepts a high-level business goal and converts it into a structured execution graph.

The system:

• decomposes the goal into specialised tasks;
• identifies dependencies between tasks;
• groups independent tasks into parallel execution waves;
• assigns specialised AI roles;
• executes tasks through GPT-5.6;
• validates task completion;
• tracks token usage, runtime, and cost;
• enforces a hard execution budget;
• consolidates the outputs into one launch package.

Instead of sending one large prompt to a model, Business OS coordinates a transparent, dependency-aware workflow.

How we built it

The MVP is implemented in Python and uses the OpenAI API as its execution layer.

The architecture includes:

• a deterministic planner;
• dependency-aware orchestration;
• parallel task waves;
• specialised role prompts;
• output validation;
• retry and failure handling;
• token and cost telemetry;
• execution journals;
• configurable budget limits;
• automated tests;
• reproducible PowerShell demo scripts.

Codex was used as the primary engineering collaborator for architecture review, implementation support, test expansion, validation, documentation, release auditing, and preparation of the public repository.

The major product and architecture decisions remained human-directed, including the separation of planning and execution, explicit dependency handling, hard cost ceilings, transparent telemetry, and human control over the final result.

Real execution result

The recorded demonstration shows a real OpenAI API execution:

• 4 completed specialised tasks;
• 2 dependency-aware execution waves;
• 22.7 seconds runtime;
• 8,327 input tokens;
• 3,876 output tokens;
• $0.0316 total execution cost;
• $0.2500 enforced cost limit;
• 50 passing automated tests.

Challenges we ran into

The main challenge was transforming an open-ended business objective into a controlled execution system instead of an unstructured sequence of prompts.

The platform needed deterministic dependency handling, parallel execution, validation, budget enforcement, error handling, and sufficient observability to explain what happened during every run.

Another challenge was producing a credible, reproducible, and safe public release within the Build Week deadline.

Accomplishments that we are proud of

• A working dependency-aware AI orchestration engine;
• real GPT-5.6 execution through the OpenAI API;
• parallel execution waves with deterministic ordering;
• hard cost limits and detailed telemetry;
• structured validation and execution journals;
• 50 passing automated tests;
• a public and reproducible GitHub repository;
• documented Codex collaboration and release audit.

What we learned

Reliable agentic systems need more than capable models. They require orchestration, budgets, validation, observability, and explicit human control.

The quality of an AI system depends not only on the generated answer, but also on how tasks are decomposed, scheduled, verified, measured, and consolidated.

What's next for Business OS

The next phase will add:

• a user discovery and business diagnosis layer;
• persistent project workspaces;
• human approval gates;
• external research and evidence tools;
• supplier and service integrations;
• document and communication automation;
• calibrated confidence scoring;
• reusable industry templates;
• multi-model routing based on quality, latency, and cost;
• a complete web interface.

The long-term goal is to make Business OS the central AI manager that helps a person discover, validate, design, and launch a business while keeping the human in control.

Codex Session / Thread ID: 019f7f74-a8db-7ee3-bd66-3c3c412b3fd0

Built With

• codex
• gpt-5.6
• openai-api
• powershell
• pytest
• python